← Paper Digest

Polygraph: Large-Scale Multi-Session Trajectory Optimization using Pose Graphs

Petar Lukovic, Tian Yi Lim, Marc Pollefeys, Pablo Speciale, Hermann Blum, Zuria Bauer

November 2026

TL;DR Polygraph merges many independently recorded camera trajectories into one consistent pose graph online, one session at a time, without building a 3D map. On a two-robot split of KITTI-00 its joint error is 1.16 m against Swarm-SLAM's 2.17 m, and it runs 1.9x faster than the closest prior method, 2GO.

What they do

Robots, vehicles and AR headsets pass through the same places many times, and each pass (a session) has its own drifting odometry in its own frame. Polygraph puts all sessions into one shared frame online, as they arrive. It fixes no reference session and builds no landmark map; its only inputs are calibrated images and odometry.

How it works

  • Clusters. Each session is thinned to keyframes with a MegaLoc global descriptor, and the keyframes are grouped into spatial clusters by travel distance.
  • Matching. A closed cluster votes for similar clusters in every other session. Keyframe pairs are verified with XFeat + LighterGlue matching, essential-matrix RANSAC, Metric3D v2 depth for metric scale and PnP RANSAC. A cluster match is confirmed only when several pair estimates agree.
  • Two-tier consensus gate. A session with no cross-session edge yet joins only when several independent, mutually consistent matches agree, never through a single edge. Later matches are checked against the current estimate or queued for a re-check.
  • Optimization. iSAM2 updates after every cluster, periodic batch solves prune outlier edges, and a final graduated non-convexity solve weights every loop closure.

Results

  • KITTI-00 split between robots: joint ATE 1.16 m vs 2.17 m for Swarm-SLAM (2 robots), and 1.26 m vs 8.00 m for DOOR-SLAM (3 robots), with RTAB-Map odometry.
  • Single session, total ATE: 6.69 m vs 9.33 m for 2GO on KITTI (ORB-SLAM3 odometry), and 26.92 m vs 41.71 m on VBR (VINS-Fusion odometry).
  • Indoor multi-session results improve every device’s own odometry, e.g. Spot mean ATE from 0.299 to 0.149 m and HoloLens from 0.658 to 0.280 m.
  • Against 2GO on an RTX 4090: 1.9x faster on KITTI-00 with 26 edges instead of 517. Without the gate, committed edges have 2.5x the translation error.

Limitations

Committed edges are less accurate than those of an IMU-fused two-view system: 0.48 m on EuRoC versus 0.07 m for COVINS-G. With ORB-SLAM3 odometry, VBR needed one config per location. The 2GO timings come from its paper rather than a rerun, and only visual and visual-inertial odometry were tested.

Figures

From Figure 1: Five independent HYDRO sessions recorded with a Boston Dynamics Spot, each in its own color, placed in...
From Figure 1: Five independent HYDRO sessions recorded with a Boston Dynamics Spot, each in its own color, placed into one shared frame by Polygraph and overlaid on the site's 3D scan. Purple lines mark confirmed cross-session loop-closure edges.
From Figure 2: Polygraph pipeline. Each session's odometry is keyframed and grouped online into spatial clusters. A c...
From Figure 2: Polygraph pipeline. Each session's odometry is keyframed and grouped online into spatial clusters. A closed cluster is matched against every other session by image retrieval, candidate pairs are verified by joint pose estimation, and a confirmed match enters the shared pose graph once the consensus gate admits it.
From Figure 6: Reference anchoring on HGE (top) and HYDRO (bottom), ground truth in green with Polygraph's estimate o...
From Figure 6: Reference anchoring on HGE (top) and HYDRO (bottom), ground truth in green with Polygraph's estimate on top. HGE: (a) HoloLens alone, split into two reference groups, (b) iOS alone, (c) HoloLens and iOS anchored together. HYDRO: (a) Spot alone, (b) Spot and iOS together, all ten sessions in one component.
From Figure 8: Runtime and committed loop-closure edges of Polygraph and 2GO on the same RTX 4090 with VINS-Fusion od...
From Figure 8: Runtime and committed loop-closure edges of Polygraph and 2GO on the same RTX 4090 with VINS-Fusion odometry, log scale. 2GO values are from its paper; Polygraph's are the mean of five runs.
Abstract

Robots, vehicles, and wearable devices often traverse the same environment many times. Each pass produces its own trajectory from onboard visual or visual-inertial odometry. These trajectories are locally accurate but drift over time and live in unrelated coordinate frames. Turning many such independent trajectories into one globally consistent set is the large-scale multi-session trajectory optimization problem. Existing solutions largely work offline. All sessions must be available up front, one session is fixed as the reference frame, and a single fused map is then optimized. This paper presents Polygraph, a multi-session trajectory optimization pipeline that processes sessions online, one keyframe and one spatial cluster at a time, and never builds a 3D map. Keyframes are grouped into local clusters. When a cluster closes, it is matched against every other session through learned image retrieval. Candidate pairs are verified with sparse feature matching, monocular metric depth, and robust pose estimation. A two-tier consensus gate admits a cross-session edge into the shared pose graph only once independent evidence corroborates it. A cheap incremental solve keeps the graph current, and a periodic robust batch solve prunes outliers. We evaluate Polygraph on stereo outdoor driving, HoloLens and iPad indoor traversals, legged indoor locomotion with a Boston Dynamics Spot, and large-scale outdoor vehicle and handheld routes. Against the closest prior work on the same GPU model, using its reported timings, Polygraph runs faster, commits fewer loop-closure edges, and matches or improves on accuracy.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.